Neopark Smart Parking System

Urban parking facilities waste enormous amounts of driver time and fuel because occupancy data is either manual or simply non-existent. Neopark set out to deliver real-time, per-slot parking availability without the cost of per-slot hardware sensors — using computer vision and low-cost edge cameras instead. To get there we built an end-to-end pipeline: ESP32-CAM edge devices stream frames to a centralized inference service running a custom-trained YOLOv11 detector, slot occupancy is computed server-side and pushed to a Next.js live dashboard, and every image and event log is archived to AWS S3, with the whole stack containerized in Docker for repeatable deployment.

The detection model was trained on a curated dataset of parking scenes under varied lighting, with confidence thresholds tuned per camera angle and a temporal smoothing layer added to suppress flicker between adjacent frames. By batching frames and running on a GPU-backed inference node we kept latency around 150ms, and the system ultimately reached 94.2% detection accuracy with 99.8% uptime across the pilot deployment. That work won the Best Project Award at the FILKOM UB Technology Exhibition 2025 and drew the most-visited booth at the event.

The biggest lesson was that edge-to-cloud reliability matters more than raw model accuracy. Investing early in temporal smoothing and per-camera calibration paid off far more than chasing the last percentage point of mAP — the difference between an impressive demo and a system people could actually trust came down to engineering, not the model alone.